I specialize in Physical AI: the convergence of Mechanical Engineering, Probabilistic Machine Learning, and Fault-Tolerant Embedded Systems. My work focuses on collapsing the distance between high-fidelity physical simulations and real-time, dependable execution on the edge.

I implement End-to-End Physical Intelligence Pipelines: High-Fidelity Sim (ANSYS) $\rightarrow$ Probabilistic ROM (MLP Ensembles) $\rightarrow$ Dependable Firmware (TMR C++/RTOS).

An industrial-grade firmware framework for ESP32 designed for safety-critical environments (e.g., flight controllers).

• Autonomous Self-Healing: Triple OTA partition management (ota_0, 1, 2) with SHA-256 validation every 60s. Automatically performs chunk-based (4KB) destructive copies to repair corrupted firmware slots.
• Two-Layer TMR: Implemented a high-order redundancy pattern for critical data and arithmetic.
  • Layer 1: Majority vote across 3 rows $\rightarrow$ 3 candidates.
  • Layer 2: Majority vote of candidates $\rightarrow$ validated result.
• Computational Redundancy: Prevents transient CPU faults by executing arithmetic operations 9 times independently with volatile barriers to prevent compiler optimization.
• Validation: Rigorous test suite that systematically injects corruption into flash memory to verify autonomous recovery.

A full-stack pipeline for the parametric optimization of MEMS accelerometer flexures.

• The Pipeline: Scripted ANSYS simulations $\rightarrow$ Ensemble Neural Network $\rightarrow$ NSGA-II Optimization.
• Probabilistic ROM: Built an ensemble of 20 MLP regressors to predict 10 structural outputs (Modal Frequencies, Stress, Deformation).
• Uncertainty Quantification: Instead of a black-box mean, the system outputs a standard deviation ($\sigma$), enabling reliability-aware design exploration.
• High Precision: Achieved $R^2$ scores up to 0.9995 for critical modal frequencies.
• Analysis: Utilized Jacobian sensitivity analysis and UMAP manifold extraction to validate the model's latent space.

A local-first AI agent built on LangGraph and Ollama, designed for autonomous engineering workflows.

• Semantic Compaction: Implemented a token-triggered compaction system that summarizes older conversation history into a rolling state, preventing context window saturation.
• Complex Tooling: Integrated a persistent shell process (pwsh), surgical str_replace for file editing, and a subagent spawning system for parallel task execution.
• Multi-Modal: Integrated vision-based OCR and PDF parsing (LlamaCloud) for high-density document processing.

I am currently refining the deployment of Physics-Informed Neural Networks (PINNs) and Reduced Order Models (ROMs) into hard real-time C++ environments. My goal is to eliminate the "Simulation Bottleneck" by enabling complex physics-based predictions to run in $<1\text{ms}$ on edge hardware.

Portfolio | LinkedIn